backend.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <limits.h>
  28 #include <signal.h>
  29 #include <time.h>
  30 #include <locale.h>
  31 #include <assert.h>
  32 #include <time.h>
  33 #include <inttypes.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #ifndef FIO_NO_HAVE_SHM_H
  38 #include <sys/shm.h>
  39 #endif
  40 #include <sys/mman.h>
  41
  42 #include "fio.h"
  43 #include "hash.h"
  44 #include "smalloc.h"
  45 #include "verify.h"
  46 #include "trim.h"
  47 #include "diskutil.h"
  48 #include "cgroup.h"
  49 #include "profile.h"
  50 #include "lib/rand.h"
  51 #include "memalign.h"
  52 #include "server.h"
  53
  54 static pthread_t disk_util_thread;
  55 static struct fio_mutex *disk_thread_mutex;
  56 static struct fio_mutex *startup_mutex;
  57 static struct fio_mutex *writeout_mutex;
  58 static struct flist_head *cgroup_list;
  59 static char *cgroup_mnt;
  60 static int exit_value;
  61 static volatile int fio_abort;
  62
  63 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
  64
  65 int groupid = 0;
  66 unsigned int thread_number = 0;
  67 unsigned int stat_number = 0;
  68 unsigned int nr_process = 0;
  69 unsigned int nr_thread = 0;
  70 int shm_id = 0;
  71 int temp_stall_ts;
  72 unsigned long done_secs = 0;
  73 volatile int disk_util_exit = 0;
  74
  75 #define PAGE_ALIGN(buf) \
  76         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
  77
  78 #define JOB_START_TIMEOUT       (5 * 1000)
  79
  80 static void sig_int(int sig)
  81 {
  82         if (threads) {
  83                 if (is_backend)
  84                         fio_server_got_signal(sig);
  85                 else {
  86                         log_info("\nfio: terminating on signal %d\n", sig);
  87                         fflush(stdout);
  88                         exit_value = 128;
  89                 }
  90
  91                 fio_terminate_threads(TERMINATE_ALL);
  92         }
  93 }
  94
  95 static void sig_show_status(int sig)
  96 {
  97         show_running_run_stats();
  98 }
  99
 100 static void set_sig_handlers(void)
 101 {
 102         struct sigaction act;
 103
 104         memset(&act, 0, sizeof(act));
 105         act.sa_handler = sig_int;
 106         act.sa_flags = SA_RESTART;
 107         sigaction(SIGINT, &act, NULL);
 108
 109         memset(&act, 0, sizeof(act));
 110         act.sa_handler = sig_int;
 111         act.sa_flags = SA_RESTART;
 112         sigaction(SIGTERM, &act, NULL);
 113
 114 /* Windows uses SIGBREAK as a quit signal from other applications */
 115 #ifdef WIN32
 116         memset(&act, 0, sizeof(act));
 117         act.sa_handler = sig_int;
 118         act.sa_flags = SA_RESTART;
 119         sigaction(SIGBREAK, &act, NULL);
 120 #endif
 121
 122         memset(&act, 0, sizeof(act));
 123         act.sa_handler = sig_show_status;
 124         act.sa_flags = SA_RESTART;
 125         sigaction(SIGUSR1, &act, NULL);
 126
 127         if (is_backend) {
 128                 memset(&act, 0, sizeof(act));
 129                 act.sa_handler = sig_int;
 130                 act.sa_flags = SA_RESTART;
 131                 sigaction(SIGPIPE, &act, NULL);
 132         }
 133 }
 134
 135 /*
 136  * Check if we are above the minimum rate given.
 137  */
 138 static int __check_min_rate(struct thread_data *td, struct timeval *now,
 139                             enum fio_ddir ddir)
 140 {
 141         unsigned long long bytes = 0;
 142         unsigned long iops = 0;
 143         unsigned long spent;
 144         unsigned long rate;
 145         unsigned int ratemin = 0;
 146         unsigned int rate_iops = 0;
 147         unsigned int rate_iops_min = 0;
 148
 149         assert(ddir_rw(ddir));
 150
 151         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
 152                 return 0;
 153
 154         /*
 155          * allow a 2 second settle period in the beginning
 156          */
 157         if (mtime_since(&td->start, now) < 2000)
 158                 return 0;
 159
 160         iops += td->this_io_blocks[ddir];
 161         bytes += td->this_io_bytes[ddir];
 162         ratemin += td->o.ratemin[ddir];
 163         rate_iops += td->o.rate_iops[ddir];
 164         rate_iops_min += td->o.rate_iops_min[ddir];
 165
 166         /*
 167          * if rate blocks is set, sample is running
 168          */
 169         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
 170                 spent = mtime_since(&td->lastrate[ddir], now);
 171                 if (spent < td->o.ratecycle)
 172                         return 0;
 173
 174                 if (td->o.rate[ddir]) {
 175                         /*
 176                          * check bandwidth specified rate
 177                          */
 178                         if (bytes < td->rate_bytes[ddir]) {
 179                                 log_err("%s: min rate %u not met\n", td->o.name,
 180                                                                 ratemin);
 181                                 return 1;
 182                         } else {
 183                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
 184                                 if (rate < ratemin ||
 185                                     bytes < td->rate_bytes[ddir]) {
 186                                         log_err("%s: min rate %u not met, got"
 187                                                 " %luKB/sec\n", td->o.name,
 188                                                         ratemin, rate);
 189                                         return 1;
 190                                 }
 191                         }
 192                 } else {
 193                         /*
 194                          * checks iops specified rate
 195                          */
 196                         if (iops < rate_iops) {
 197                                 log_err("%s: min iops rate %u not met\n",
 198                                                 td->o.name, rate_iops);
 199                                 return 1;
 200                         } else {
 201                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
 202                                 if (rate < rate_iops_min ||
 203                                     iops < td->rate_blocks[ddir]) {
 204                                         log_err("%s: min iops rate %u not met,"
 205                                                 " got %lu\n", td->o.name,
 206                                                         rate_iops_min, rate);
 207                                 }
 208                         }
 209                 }
 210         }
 211
 212         td->rate_bytes[ddir] = bytes;
 213         td->rate_blocks[ddir] = iops;
 214         memcpy(&td->lastrate[ddir], now, sizeof(*now));
 215         return 0;
 216 }
 217
 218 static int check_min_rate(struct thread_data *td, struct timeval *now,
 219                           unsigned long *bytes_done)
 220 {
 221         int ret = 0;
 222
 223         if (bytes_done[DDIR_READ])
 224                 ret |= __check_min_rate(td, now, DDIR_READ);
 225         if (bytes_done[DDIR_WRITE])
 226                 ret |= __check_min_rate(td, now, DDIR_WRITE);
 227         if (bytes_done[DDIR_TRIM])
 228                 ret |= __check_min_rate(td, now, DDIR_TRIM);
 229
 230         return ret;
 231 }
 232
 233 /*
 234  * When job exits, we can cancel the in-flight IO if we are using async
 235  * io. Attempt to do so.
 236  */
 237 static void cleanup_pending_aio(struct thread_data *td)
 238 {
 239         struct flist_head *entry, *n;
 240         struct io_u *io_u;
 241         int r;
 242
 243         /*
 244          * get immediately available events, if any
 245          */
 246         r = io_u_queued_complete(td, 0, NULL);
 247         if (r < 0)
 248                 return;
 249
 250         /*
 251          * now cancel remaining active events
 252          */
 253         if (td->io_ops->cancel) {
 254                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 255                         io_u = flist_entry(entry, struct io_u, list);
 256
 257                         /*
 258                          * if the io_u isn't in flight, then that generally
 259                          * means someone leaked an io_u. complain but fix
 260                          * it up, so we don't stall here.
 261                          */
 262                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 263                                 log_err("fio: non-busy IO on busy list\n");
 264                                 put_io_u(td, io_u);
 265                         } else {
 266                                 r = td->io_ops->cancel(td, io_u);
 267                                 if (!r)
 268                                         put_io_u(td, io_u);
 269                         }
 270                 }
 271         }
 272
 273         if (td->cur_depth)
 274                 r = io_u_queued_complete(td, td->cur_depth, NULL);
 275 }
 276
 277 /*
 278  * Helper to handle the final sync of a file. Works just like the normal
 279  * io path, just does everything sync.
 280  */
 281 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 282 {
 283         struct io_u *io_u = __get_io_u(td);
 284         int ret;
 285
 286         if (!io_u)
 287                 return 1;
 288
 289         io_u->ddir = DDIR_SYNC;
 290         io_u->file = f;
 291
 292         if (td_io_prep(td, io_u)) {
 293                 put_io_u(td, io_u);
 294                 return 1;
 295         }
 296
 297 requeue:
 298         ret = td_io_queue(td, io_u);
 299         if (ret < 0) {
 300                 td_verror(td, io_u->error, "td_io_queue");
 301                 put_io_u(td, io_u);
 302                 return 1;
 303         } else if (ret == FIO_Q_QUEUED) {
 304                 if (io_u_queued_complete(td, 1, NULL) < 0)
 305                         return 1;
 306         } else if (ret == FIO_Q_COMPLETED) {
 307                 if (io_u->error) {
 308                         td_verror(td, io_u->error, "td_io_queue");
 309                         return 1;
 310                 }
 311
 312                 if (io_u_sync_complete(td, io_u, NULL) < 0)
 313                         return 1;
 314         } else if (ret == FIO_Q_BUSY) {
 315                 if (td_io_commit(td))
 316                         return 1;
 317                 goto requeue;
 318         }
 319
 320         return 0;
 321 }
 322
 323 static inline void __update_tv_cache(struct thread_data *td)
 324 {
 325         fio_gettime(&td->tv_cache, NULL);
 326 }
 327
 328 static inline void update_tv_cache(struct thread_data *td)
 329 {
 330         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
 331                 __update_tv_cache(td);
 332 }
 333
 334 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 335 {
 336         if (in_ramp_time(td))
 337                 return 0;
 338         if (!td->o.timeout)
 339                 return 0;
 340         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 341                 return 1;
 342
 343         return 0;
 344 }
 345
 346 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
 347                                int *retptr)
 348 {
 349         int ret = *retptr;
 350
 351         if (ret < 0 || td->error) {
 352                 int err = td->error;
 353                 enum error_type_bit eb;
 354
 355                 if (ret < 0)
 356                         err = -ret;
 357
 358                 eb = td_error_type(ddir, err);
 359                 if (!(td->o.continue_on_error & (1 << eb)))
 360                         return 1;
 361
 362                 if (td_non_fatal_error(td, eb, err)) {
 363                         /*
 364                          * Continue with the I/Os in case of
 365                          * a non fatal error.
 366                          */
 367                         update_error_count(td, err);
 368                         td_clear_error(td);
 369                         *retptr = 0;
 370                         return 0;
 371                 } else if (td->o.fill_device && err == ENOSPC) {
 372                         /*
 373                          * We expect to hit this error if
 374                          * fill_device option is set.
 375                          */
 376                         td_clear_error(td);
 377                         td->terminate = 1;
 378                         return 1;
 379                 } else {
 380                         /*
 381                          * Stop the I/O in case of a fatal
 382                          * error.
 383                          */
 384                         update_error_count(td, err);
 385                         return 1;
 386                 }
 387         }
 388
 389         return 0;
 390 }
 391
 392 /*
 393  * The main verify engine. Runs over the writes we previously submitted,
 394  * reads the blocks back in, and checks the crc/md5 of the data.
 395  */
 396 static void do_verify(struct thread_data *td)
 397 {
 398         struct fio_file *f;
 399         struct io_u *io_u;
 400         int ret, min_events;
 401         unsigned int i;
 402
 403         dprint(FD_VERIFY, "starting loop\n");
 404
 405         /*
 406          * sync io first and invalidate cache, to make sure we really
 407          * read from disk.
 408          */
 409         for_each_file(td, f, i) {
 410                 if (!fio_file_open(f))
 411                         continue;
 412                 if (fio_io_sync(td, f))
 413                         break;
 414                 if (file_invalidate_cache(td, f))
 415                         break;
 416         }
 417
 418         if (td->error)
 419                 return;
 420
 421         td_set_runstate(td, TD_VERIFYING);
 422
 423         io_u = NULL;
 424         while (!td->terminate) {
 425                 enum fio_ddir ddir;
 426                 int ret2, full;
 427
 428                 update_tv_cache(td);
 429
 430                 if (runtime_exceeded(td, &td->tv_cache)) {
 431                         __update_tv_cache(td);
 432                         if (runtime_exceeded(td, &td->tv_cache)) {
 433                                 td->terminate = 1;
 434                                 break;
 435                         }
 436                 }
 437
 438                 if (flow_threshold_exceeded(td))
 439                         continue;
 440
 441                 if (!td->o.experimental_verify) {
 442                         io_u = __get_io_u(td);
 443                         if (!io_u)
 444                                 break;
 445
 446                         if (get_next_verify(td, io_u)) {
 447                                 put_io_u(td, io_u);
 448                                 break;
 449                         }
 450
 451                         if (td_io_prep(td, io_u)) {
 452                                 put_io_u(td, io_u);
 453                                 break;
 454                         }
 455                 } else {
 456                         while ((io_u = get_io_u(td)) != NULL) {
 457                                 /*
 458                                  * We are only interested in the places where
 459                                  * we wrote or trimmed IOs. Turn those into
 460                                  * reads for verification purposes.
 461                                  */
 462                                 if (io_u->ddir == DDIR_READ) {
 463                                         /*
 464                                          * Pretend we issued it for rwmix
 465                                          * accounting
 466                                          */
 467                                         td->io_issues[DDIR_READ]++;
 468                                         put_io_u(td, io_u);
 469                                         continue;
 470                                 } else if (io_u->ddir == DDIR_TRIM) {
 471                                         io_u->ddir = DDIR_READ;
 472                                         io_u->flags |= IO_U_F_TRIMMED;
 473                                         break;
 474                                 } else if (io_u->ddir == DDIR_WRITE) {
 475                                         io_u->ddir = DDIR_READ;
 476                                         break;
 477                                 } else {
 478                                         put_io_u(td, io_u);
 479                                         continue;
 480                                 }
 481                         }
 482
 483                         if (!io_u)
 484                                 break;
 485                 }
 486
 487                 if (td->o.verify_async)
 488                         io_u->end_io = verify_io_u_async;
 489                 else
 490                         io_u->end_io = verify_io_u;
 491
 492                 ddir = io_u->ddir;
 493
 494                 ret = td_io_queue(td, io_u);
 495                 switch (ret) {
 496                 case FIO_Q_COMPLETED:
 497                         if (io_u->error) {
 498                                 ret = -io_u->error;
 499                                 clear_io_u(td, io_u);
 500                         } else if (io_u->resid) {
 501                                 int bytes = io_u->xfer_buflen - io_u->resid;
 502
 503                                 /*
 504                                  * zero read, fail
 505                                  */
 506                                 if (!bytes) {
 507                                         td_verror(td, EIO, "full resid");
 508                                         put_io_u(td, io_u);
 509                                         break;
 510                                 }
 511
 512                                 io_u->xfer_buflen = io_u->resid;
 513                                 io_u->xfer_buf += bytes;
 514                                 io_u->offset += bytes;
 515
 516                                 if (ddir_rw(io_u->ddir))
 517                                         td->ts.short_io_u[io_u->ddir]++;
 518
 519                                 f = io_u->file;
 520                                 if (io_u->offset == f->real_file_size)
 521                                         goto sync_done;
 522
 523                                 requeue_io_u(td, &io_u);
 524                         } else {
 525 sync_done:
 526                                 ret = io_u_sync_complete(td, io_u, NULL);
 527                                 if (ret < 0)
 528                                         break;
 529                         }
 530                         continue;
 531                 case FIO_Q_QUEUED:
 532                         break;
 533                 case FIO_Q_BUSY:
 534                         requeue_io_u(td, &io_u);
 535                         ret2 = td_io_commit(td);
 536                         if (ret2 < 0)
 537                                 ret = ret2;
 538                         break;
 539                 default:
 540                         assert(ret < 0);
 541                         td_verror(td, -ret, "td_io_queue");
 542                         break;
 543                 }
 544
 545                 if (break_on_this_error(td, ddir, &ret))
 546                         break;
 547
 548                 /*
 549                  * if we can queue more, do so. but check if there are
 550                  * completed io_u's first. Note that we can get BUSY even
 551                  * without IO queued, if the system is resource starved.
 552                  */
 553                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 554                 if (full || !td->o.iodepth_batch_complete) {
 555                         min_events = min(td->o.iodepth_batch_complete,
 556                                          td->cur_depth);
 557                         /*
 558                          * if the queue is full, we MUST reap at least 1 event
 559                          */
 560                         if (full && !min_events)
 561                                 min_events = 1;
 562
 563                         do {
 564                                 /*
 565                                  * Reap required number of io units, if any,
 566                                  * and do the verification on them through
 567                                  * the callback handler
 568                                  */
 569                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
 570                                         ret = -1;
 571                                         break;
 572                                 }
 573                         } while (full && (td->cur_depth > td->o.iodepth_low));
 574                 }
 575                 if (ret < 0)
 576                         break;
 577         }
 578
 579         if (!td->error) {
 580                 min_events = td->cur_depth;
 581
 582                 if (min_events)
 583                         ret = io_u_queued_complete(td, min_events, NULL);
 584         } else
 585                 cleanup_pending_aio(td);
 586
 587         td_set_runstate(td, TD_RUNNING);
 588
 589         dprint(FD_VERIFY, "exiting loop\n");
 590 }
 591
 592 static int io_bytes_exceeded(struct thread_data *td)
 593 {
 594         unsigned long long bytes;
 595
 596         if (td_rw(td))
 597                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
 598         else if (td_write(td))
 599                 bytes = td->this_io_bytes[DDIR_WRITE];
 600         else if (td_read(td))
 601                 bytes = td->this_io_bytes[DDIR_READ];
 602         else
 603                 bytes = td->this_io_bytes[DDIR_TRIM];
 604
 605         return bytes >= td->o.size;
 606 }
 607
 608 /*
 609  * Main IO worker function. It retrieves io_u's to process and queues
 610  * and reaps them, checking for rate and errors along the way.
 611  */
 612 static void do_io(struct thread_data *td)
 613 {
 614         unsigned int i;
 615         int ret = 0;
 616
 617         if (in_ramp_time(td))
 618                 td_set_runstate(td, TD_RAMP);
 619         else
 620                 td_set_runstate(td, TD_RUNNING);
 621
 622         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
 623                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
 624                 td->o.time_based) {
 625                 struct timeval comp_time;
 626                 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
 627                 int min_evts = 0;
 628                 struct io_u *io_u;
 629                 int ret2, full;
 630                 enum fio_ddir ddir;
 631
 632                 if (td->terminate || td->done)
 633                         break;
 634
 635                 update_tv_cache(td);
 636
 637                 if (runtime_exceeded(td, &td->tv_cache)) {
 638                         __update_tv_cache(td);
 639                         if (runtime_exceeded(td, &td->tv_cache)) {
 640                                 td->terminate = 1;
 641                                 break;
 642                         }
 643                 }
 644
 645                 if (flow_threshold_exceeded(td))
 646                         continue;
 647
 648                 io_u = get_io_u(td);
 649                 if (!io_u)
 650                         break;
 651
 652                 ddir = io_u->ddir;
 653
 654                 /*
 655                  * Add verification end_io handler if:
 656                  *      - Asked to verify (!td_rw(td))
 657                  *      - Or the io_u is from our verify list (mixed write/ver)
 658                  */
 659                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
 660                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 661                         if (td->o.verify_async)
 662                                 io_u->end_io = verify_io_u_async;
 663                         else
 664                                 io_u->end_io = verify_io_u;
 665                         td_set_runstate(td, TD_VERIFYING);
 666                 } else if (in_ramp_time(td))
 667                         td_set_runstate(td, TD_RAMP);
 668                 else
 669                         td_set_runstate(td, TD_RUNNING);
 670
 671                 ret = td_io_queue(td, io_u);
 672                 switch (ret) {
 673                 case FIO_Q_COMPLETED:
 674                         if (io_u->error) {
 675                                 ret = -io_u->error;
 676                                 clear_io_u(td, io_u);
 677                         } else if (io_u->resid) {
 678                                 int bytes = io_u->xfer_buflen - io_u->resid;
 679                                 struct fio_file *f = io_u->file;
 680
 681                                 /*
 682                                  * zero read, fail
 683                                  */
 684                                 if (!bytes) {
 685                                         td_verror(td, EIO, "full resid");
 686                                         put_io_u(td, io_u);
 687                                         break;
 688                                 }
 689
 690                                 io_u->xfer_buflen = io_u->resid;
 691                                 io_u->xfer_buf += bytes;
 692                                 io_u->offset += bytes;
 693
 694                                 if (ddir_rw(io_u->ddir))
 695                                         td->ts.short_io_u[io_u->ddir]++;
 696
 697                                 if (io_u->offset == f->real_file_size)
 698                                         goto sync_done;
 699
 700                                 requeue_io_u(td, &io_u);
 701                         } else {
 702 sync_done:
 703                                 if (__should_check_rate(td, DDIR_READ) ||
 704                                     __should_check_rate(td, DDIR_WRITE) ||
 705                                     __should_check_rate(td, DDIR_TRIM))
 706                                         fio_gettime(&comp_time, NULL);
 707
 708                                 ret = io_u_sync_complete(td, io_u, bytes_done);
 709                                 if (ret < 0)
 710                                         break;
 711                         }
 712                         break;
 713                 case FIO_Q_QUEUED:
 714                         /*
 715                          * if the engine doesn't have a commit hook,
 716                          * the io_u is really queued. if it does have such
 717                          * a hook, it has to call io_u_queued() itself.
 718                          */
 719                         if (td->io_ops->commit == NULL)
 720                                 io_u_queued(td, io_u);
 721                         break;
 722                 case FIO_Q_BUSY:
 723                         requeue_io_u(td, &io_u);
 724                         ret2 = td_io_commit(td);
 725                         if (ret2 < 0)
 726                                 ret = ret2;
 727                         break;
 728                 default:
 729                         assert(ret < 0);
 730                         put_io_u(td, io_u);
 731                         break;
 732                 }
 733
 734                 if (break_on_this_error(td, ddir, &ret))
 735                         break;
 736
 737                 /*
 738                  * See if we need to complete some commands. Note that we
 739                  * can get BUSY even without IO queued, if the system is
 740                  * resource starved.
 741                  */
 742                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
 743                 if (full || !td->o.iodepth_batch_complete) {
 744                         min_evts = min(td->o.iodepth_batch_complete,
 745                                         td->cur_depth);
 746                         /*
 747                          * if the queue is full, we MUST reap at least 1 event
 748                          */
 749                         if (full && !min_evts)
 750                                 min_evts = 1;
 751
 752                         if (__should_check_rate(td, DDIR_READ) ||
 753                             __should_check_rate(td, DDIR_WRITE) ||
 754                             __should_check_rate(td, DDIR_TRIM))
 755                                 fio_gettime(&comp_time, NULL);
 756
 757                         do {
 758                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
 759                                 if (ret < 0)
 760                                         break;
 761
 762                         } while (full && (td->cur_depth > td->o.iodepth_low));
 763                 }
 764
 765                 if (ret < 0)
 766                         break;
 767                 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
 768                         continue;
 769
 770                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
 771                         if (check_min_rate(td, &comp_time, bytes_done)) {
 772                                 if (exitall_on_terminate)
 773                                         fio_terminate_threads(td->groupid);
 774                                 td_verror(td, EIO, "check_min_rate");
 775                                 break;
 776                         }
 777                 }
 778
 779                 if (td->o.thinktime) {
 780                         unsigned long long b;
 781
 782                         b = ddir_rw_sum(td->io_blocks);
 783                         if (!(b % td->o.thinktime_blocks)) {
 784                                 int left;
 785
 786                                 if (td->o.thinktime_spin)
 787                                         usec_spin(td->o.thinktime_spin);
 788
 789                                 left = td->o.thinktime - td->o.thinktime_spin;
 790                                 if (left)
 791                                         usec_sleep(td, left);
 792                         }
 793                 }
 794         }
 795
 796         if (td->trim_entries)
 797                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 798
 799         if (td->o.fill_device && td->error == ENOSPC) {
 800                 td->error = 0;
 801                 td->terminate = 1;
 802         }
 803         if (!td->error) {
 804                 struct fio_file *f;
 805
 806                 i = td->cur_depth;
 807                 if (i) {
 808                         ret = io_u_queued_complete(td, i, NULL);
 809                         if (td->o.fill_device && td->error == ENOSPC)
 810                                 td->error = 0;
 811                 }
 812
 813                 if (should_fsync(td) && td->o.end_fsync) {
 814                         td_set_runstate(td, TD_FSYNCING);
 815
 816                         for_each_file(td, f, i) {
 817                                 if (!fio_file_open(f))
 818                                         continue;
 819                                 fio_io_sync(td, f);
 820                         }
 821                 }
 822         } else
 823                 cleanup_pending_aio(td);
 824
 825         /*
 826          * stop job if we failed doing any IO
 827          */
 828         if (!ddir_rw_sum(td->this_io_bytes))
 829                 td->done = 1;
 830 }
 831
 832 static void cleanup_io_u(struct thread_data *td)
 833 {
 834         struct flist_head *entry, *n;
 835         struct io_u *io_u;
 836
 837         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 838                 io_u = flist_entry(entry, struct io_u, list);
 839
 840                 flist_del(&io_u->list);
 841
 842                 if (td->io_ops->io_u_free)
 843                         td->io_ops->io_u_free(td, io_u);
 844
 845                 fio_memfree(io_u, sizeof(*io_u));
 846         }
 847
 848         free_io_mem(td);
 849 }
 850
 851 static int init_io_u(struct thread_data *td)
 852 {
 853         struct io_u *io_u;
 854         unsigned int max_bs, min_write;
 855         int cl_align, i, max_units;
 856         int data_xfer = 1;
 857         char *p;
 858
 859         max_units = td->o.iodepth;
 860         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 861         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
 862         min_write = td->o.min_bs[DDIR_WRITE];
 863         td->orig_buffer_size = (unsigned long long) max_bs
 864                                         * (unsigned long long) max_units;
 865
 866         if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
 867                 data_xfer = 0;
 868
 869         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 870                 unsigned long bs;
 871
 872                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 873                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 874         }
 875
 876         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 877                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 878                 return 1;
 879         }
 880
 881         if (data_xfer && allocate_io_mem(td))
 882                 return 1;
 883
 884         if (td->o.odirect || td->o.mem_align ||
 885             (td->io_ops->flags & FIO_RAWIO))
 886                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
 887         else
 888                 p = td->orig_buffer;
 889
 890         cl_align = os_cache_line_size();
 891
 892         for (i = 0; i < max_units; i++) {
 893                 void *ptr;
 894
 895                 if (td->terminate)
 896                         return 1;
 897
 898                 ptr = fio_memalign(cl_align, sizeof(*io_u));
 899                 if (!ptr) {
 900                         log_err("fio: unable to allocate aligned memory\n");
 901                         break;
 902                 }
 903
 904                 io_u = ptr;
 905                 memset(io_u, 0, sizeof(*io_u));
 906                 INIT_FLIST_HEAD(&io_u->list);
 907                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 908
 909                 if (data_xfer) {
 910                         io_u->buf = p;
 911                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 912
 913                         if (td_write(td))
 914                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
 915                         if (td_write(td) && td->o.verify_pattern_bytes) {
 916                                 /*
 917                                  * Fill the buffer with the pattern if we are
 918                                  * going to be doing writes.
 919                                  */
 920                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
 921                         }
 922                 }
 923
 924                 io_u->index = i;
 925                 io_u->flags = IO_U_F_FREE;
 926                 flist_add(&io_u->list, &td->io_u_freelist);
 927
 928                 if (td->io_ops->io_u_init) {
 929                         int ret = td->io_ops->io_u_init(td, io_u);
 930
 931                         if (ret) {
 932                                 log_err("fio: failed to init engine data: %d\n", ret);
 933                                 return 1;
 934                         }
 935                 }
 936
 937                 p += max_bs;
 938         }
 939
 940         return 0;
 941 }
 942
 943 static int switch_ioscheduler(struct thread_data *td)
 944 {
 945         char tmp[256], tmp2[128];
 946         FILE *f;
 947         int ret;
 948
 949         if (td->io_ops->flags & FIO_DISKLESSIO)
 950                 return 0;
 951
 952         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 953
 954         f = fopen(tmp, "r+");
 955         if (!f) {
 956                 if (errno == ENOENT) {
 957                         log_err("fio: os or kernel doesn't support IO scheduler"
 958                                 " switching\n");
 959                         return 0;
 960                 }
 961                 td_verror(td, errno, "fopen iosched");
 962                 return 1;
 963         }
 964
 965         /*
 966          * Set io scheduler.
 967          */
 968         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 969         if (ferror(f) || ret != 1) {
 970                 td_verror(td, errno, "fwrite");
 971                 fclose(f);
 972                 return 1;
 973         }
 974
 975         rewind(f);
 976
 977         /*
 978          * Read back and check that the selected scheduler is now the default.
 979          */
 980         ret = fread(tmp, 1, sizeof(tmp), f);
 981         if (ferror(f) || ret < 0) {
 982                 td_verror(td, errno, "fread");
 983                 fclose(f);
 984                 return 1;
 985         }
 986
 987         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 988         if (!strstr(tmp, tmp2)) {
 989                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 990                 td_verror(td, EINVAL, "iosched_switch");
 991                 fclose(f);
 992                 return 1;
 993         }
 994
 995         fclose(f);
 996         return 0;
 997 }
 998
 999 static int keep_running(struct thread_data *td)
1000 {
1001         if (td->done)
1002                 return 0;
1003         if (td->o.time_based)
1004                 return 1;
1005         if (td->o.loops) {
1006                 td->o.loops--;
1007                 return 1;
1008         }
1009
1010         if (ddir_rw_sum(td->io_bytes) < td->o.size)
1011                 return 1;
1012
1013         return 0;
1014 }
1015
1016 static int exec_string(const char *string)
1017 {
1018         int ret, newlen = strlen(string) + 1 + 8;
1019         char *str;
1020
1021         str = malloc(newlen);
1022         sprintf(str, "sh -c %s", string);
1023
1024         ret = system(str);
1025         if (ret == -1)
1026                 log_err("fio: exec of cmd <%s> failed\n", str);
1027
1028         free(str);
1029         return ret;
1030 }
1031
1032 /*
1033  * Entry point for the thread based jobs. The process based jobs end up
1034  * here as well, after a little setup.
1035  */
1036 static void *thread_main(void *data)
1037 {
1038         unsigned long long elapsed;
1039         struct thread_data *td = data;
1040         pthread_condattr_t attr;
1041         int clear_state;
1042
1043         if (!td->o.use_thread) {
1044                 setsid();
1045                 td->pid = getpid();
1046         } else
1047                 td->pid = gettid();
1048
1049         fio_local_clock_init(td->o.use_thread);
1050
1051         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1052
1053         INIT_FLIST_HEAD(&td->io_u_freelist);
1054         INIT_FLIST_HEAD(&td->io_u_busylist);
1055         INIT_FLIST_HEAD(&td->io_u_requeues);
1056         INIT_FLIST_HEAD(&td->io_log_list);
1057         INIT_FLIST_HEAD(&td->io_hist_list);
1058         INIT_FLIST_HEAD(&td->verify_list);
1059         INIT_FLIST_HEAD(&td->trim_list);
1060         INIT_FLIST_HEAD(&td->next_rand_list);
1061         pthread_mutex_init(&td->io_u_lock, NULL);
1062         td->io_hist_tree = RB_ROOT;
1063
1064         pthread_condattr_init(&attr);
1065         pthread_cond_init(&td->verify_cond, &attr);
1066         pthread_cond_init(&td->free_cond, &attr);
1067
1068         td_set_runstate(td, TD_INITIALIZED);
1069         dprint(FD_MUTEX, "up startup_mutex\n");
1070         fio_mutex_up(startup_mutex);
1071         dprint(FD_MUTEX, "wait on td->mutex\n");
1072         fio_mutex_down(td->mutex);
1073         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1074
1075         /*
1076          * the ->mutex mutex is now no longer used, close it to avoid
1077          * eating a file descriptor
1078          */
1079         fio_mutex_remove(td->mutex);
1080
1081         /*
1082          * A new gid requires privilege, so we need to do this before setting
1083          * the uid.
1084          */
1085         if (td->o.gid != -1U && setgid(td->o.gid)) {
1086                 td_verror(td, errno, "setgid");
1087                 goto err;
1088         }
1089         if (td->o.uid != -1U && setuid(td->o.uid)) {
1090                 td_verror(td, errno, "setuid");
1091                 goto err;
1092         }
1093
1094         /*
1095          * If we have a gettimeofday() thread, make sure we exclude that
1096          * thread from this job
1097          */
1098         if (td->o.gtod_cpu)
1099                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1100
1101         /*
1102          * Set affinity first, in case it has an impact on the memory
1103          * allocations.
1104          */
1105         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1106                 td_verror(td, errno, "cpu_set_affinity");
1107                 goto err;
1108         }
1109
1110 #ifdef CONFIG_LIBNUMA
1111         /* numa node setup */
1112         if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1113                 int ret;
1114
1115                 if (numa_available() < 0) {
1116                         td_verror(td, errno, "Does not support NUMA API\n");
1117                         goto err;
1118                 }
1119
1120                 if (td->o.numa_cpumask_set) {
1121                         ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1122                         if (ret == -1) {
1123                                 td_verror(td, errno, \
1124                                         "numa_run_on_node_mask failed\n");
1125                                 goto err;
1126                         }
1127                 }
1128
1129                 if (td->o.numa_memmask_set) {
1130
1131                         switch (td->o.numa_mem_mode) {
1132                         case MPOL_INTERLEAVE:
1133                                 numa_set_interleave_mask(td->o.numa_memnodesmask);
1134                                 break;
1135                         case MPOL_BIND:
1136                                 numa_set_membind(td->o.numa_memnodesmask);
1137                                 break;
1138                         case MPOL_LOCAL:
1139                                 numa_set_localalloc();
1140                                 break;
1141                         case MPOL_PREFERRED:
1142                                 numa_set_preferred(td->o.numa_mem_prefer_node);
1143                                 break;
1144                         case MPOL_DEFAULT:
1145                         default:
1146                                 break;
1147                         }
1148
1149                 }
1150         }
1151 #endif
1152
1153         /*
1154          * May alter parameters that init_io_u() will use, so we need to
1155          * do this first.
1156          */
1157         if (init_iolog(td))
1158                 goto err;
1159
1160         if (init_io_u(td))
1161                 goto err;
1162
1163         if (td->o.verify_async && verify_async_init(td))
1164                 goto err;
1165
1166         if (td->ioprio_set) {
1167                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1168                         td_verror(td, errno, "ioprio_set");
1169                         goto err;
1170                 }
1171         }
1172
1173         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1174                 goto err;
1175
1176         errno = 0;
1177         if (nice(td->o.nice) == -1 && errno != 0) {
1178                 td_verror(td, errno, "nice");
1179                 goto err;
1180         }
1181
1182         if (td->o.ioscheduler && switch_ioscheduler(td))
1183                 goto err;
1184
1185         if (!td->o.create_serialize && setup_files(td))
1186                 goto err;
1187
1188         if (td_io_init(td))
1189                 goto err;
1190
1191         if (init_random_map(td))
1192                 goto err;
1193
1194         if (td->o.exec_prerun) {
1195                 if (exec_string(td->o.exec_prerun))
1196                         goto err;
1197         }
1198
1199         if (td->o.pre_read) {
1200                 if (pre_read_files(td) < 0)
1201                         goto err;
1202         }
1203
1204         fio_gettime(&td->epoch, NULL);
1205         getrusage(RUSAGE_SELF, &td->ru_start);
1206
1207         clear_state = 0;
1208         while (keep_running(td)) {
1209                 fio_gettime(&td->start, NULL);
1210                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1211                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1212                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1213
1214                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1215                                 td->o.ratemin[DDIR_TRIM]) {
1216                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1217                                                 sizeof(td->bw_sample_time));
1218                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1219                                                 sizeof(td->bw_sample_time));
1220                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1221                                                 sizeof(td->bw_sample_time));
1222                 }
1223
1224                 if (clear_state)
1225                         clear_io_state(td);
1226
1227                 prune_io_piece_log(td);
1228
1229                 do_io(td);
1230
1231                 clear_state = 1;
1232
1233                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1234                         elapsed = utime_since_now(&td->start);
1235                         td->ts.runtime[DDIR_READ] += elapsed;
1236                 }
1237                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1238                         elapsed = utime_since_now(&td->start);
1239                         td->ts.runtime[DDIR_WRITE] += elapsed;
1240                 }
1241                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1242                         elapsed = utime_since_now(&td->start);
1243                         td->ts.runtime[DDIR_TRIM] += elapsed;
1244                 }
1245
1246                 if (td->error || td->terminate)
1247                         break;
1248
1249                 if (!td->o.do_verify ||
1250                     td->o.verify == VERIFY_NONE ||
1251                     (td->io_ops->flags & FIO_UNIDIR))
1252                         continue;
1253
1254                 clear_io_state(td);
1255
1256                 fio_gettime(&td->start, NULL);
1257
1258                 do_verify(td);
1259
1260                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1261
1262                 if (td->error || td->terminate)
1263                         break;
1264         }
1265
1266         update_rusage_stat(td);
1267         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1268         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1269         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1270         td->ts.total_run_time = mtime_since_now(&td->epoch);
1271         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1272         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1273         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1274
1275         fio_mutex_down(writeout_mutex);
1276         if (td->bw_log) {
1277                 if (td->o.bw_log_file) {
1278                         finish_log_named(td, td->bw_log,
1279                                                 td->o.bw_log_file, "bw");
1280                 } else
1281                         finish_log(td, td->bw_log, "bw");
1282         }
1283         if (td->lat_log) {
1284                 if (td->o.lat_log_file) {
1285                         finish_log_named(td, td->lat_log,
1286                                                 td->o.lat_log_file, "lat");
1287                 } else
1288                         finish_log(td, td->lat_log, "lat");
1289         }
1290         if (td->slat_log) {
1291                 if (td->o.lat_log_file) {
1292                         finish_log_named(td, td->slat_log,
1293                                                 td->o.lat_log_file, "slat");
1294                 } else
1295                         finish_log(td, td->slat_log, "slat");
1296         }
1297         if (td->clat_log) {
1298                 if (td->o.lat_log_file) {
1299                         finish_log_named(td, td->clat_log,
1300                                                 td->o.lat_log_file, "clat");
1301                 } else
1302                         finish_log(td, td->clat_log, "clat");
1303         }
1304         if (td->iops_log) {
1305                 if (td->o.iops_log_file) {
1306                         finish_log_named(td, td->iops_log,
1307                                                 td->o.iops_log_file, "iops");
1308                 } else
1309                         finish_log(td, td->iops_log, "iops");
1310         }
1311
1312         fio_mutex_up(writeout_mutex);
1313         if (td->o.exec_postrun)
1314                 exec_string(td->o.exec_postrun);
1315
1316         if (exitall_on_terminate)
1317                 fio_terminate_threads(td->groupid);
1318
1319 err:
1320         if (td->error)
1321                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1322                                                         td->verror);
1323
1324         if (td->o.verify_async)
1325                 verify_async_exit(td);
1326
1327         close_and_free_files(td);
1328         cleanup_io_u(td);
1329         close_ioengine(td);
1330         cgroup_shutdown(td, &cgroup_mnt);
1331
1332         if (td->o.cpumask_set) {
1333                 int ret = fio_cpuset_exit(&td->o.cpumask);
1334
1335                 td_verror(td, ret, "fio_cpuset_exit");
1336         }
1337
1338         /*
1339          * do this very late, it will log file closing as well
1340          */
1341         if (td->o.write_iolog_file)
1342                 write_iolog_close(td);
1343
1344         td_set_runstate(td, TD_EXITED);
1345         return (void *) (uintptr_t) td->error;
1346 }
1347
1348
1349 /*
1350  * We cannot pass the td data into a forked process, so attach the td and
1351  * pass it to the thread worker.
1352  */
1353 static int fork_main(int shmid, int offset)
1354 {
1355         struct thread_data *td;
1356         void *data, *ret;
1357
1358 #ifndef __hpux
1359         data = shmat(shmid, NULL, 0);
1360         if (data == (void *) -1) {
1361                 int __err = errno;
1362
1363                 perror("shmat");
1364                 return __err;
1365         }
1366 #else
1367         /*
1368          * HP-UX inherits shm mappings?
1369          */
1370         data = threads;
1371 #endif
1372
1373         td = data + offset * sizeof(struct thread_data);
1374         ret = thread_main(td);
1375         shmdt(data);
1376         return (int) (uintptr_t) ret;
1377 }
1378
1379 /*
1380  * Run over the job map and reap the threads that have exited, if any.
1381  */
1382 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1383                          unsigned int *m_rate)
1384 {
1385         struct thread_data *td;
1386         unsigned int cputhreads, realthreads, pending;
1387         int i, status, ret;
1388
1389         /*
1390          * reap exited threads (TD_EXITED -> TD_REAPED)
1391          */
1392         realthreads = pending = cputhreads = 0;
1393         for_each_td(td, i) {
1394                 int flags = 0;
1395
1396                 /*
1397                  * ->io_ops is NULL for a thread that has closed its
1398                  * io engine
1399                  */
1400                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1401                         cputhreads++;
1402                 else
1403                         realthreads++;
1404
1405                 if (!td->pid) {
1406                         pending++;
1407                         continue;
1408                 }
1409                 if (td->runstate == TD_REAPED)
1410                         continue;
1411                 if (td->o.use_thread) {
1412                         if (td->runstate == TD_EXITED) {
1413                                 td_set_runstate(td, TD_REAPED);
1414                                 goto reaped;
1415                         }
1416                         continue;
1417                 }
1418
1419                 flags = WNOHANG;
1420                 if (td->runstate == TD_EXITED)
1421                         flags = 0;
1422
1423                 /*
1424                  * check if someone quit or got killed in an unusual way
1425                  */
1426                 ret = waitpid(td->pid, &status, flags);
1427                 if (ret < 0) {
1428                         if (errno == ECHILD) {
1429                                 log_err("fio: pid=%d disappeared %d\n",
1430                                                 (int) td->pid, td->runstate);
1431                                 td->sig = ECHILD;
1432                                 td_set_runstate(td, TD_REAPED);
1433                                 goto reaped;
1434                         }
1435                         perror("waitpid");
1436                 } else if (ret == td->pid) {
1437                         if (WIFSIGNALED(status)) {
1438                                 int sig = WTERMSIG(status);
1439
1440                                 if (sig != SIGTERM && sig != SIGUSR2)
1441                                         log_err("fio: pid=%d, got signal=%d\n",
1442                                                         (int) td->pid, sig);
1443                                 td->sig = sig;
1444                                 td_set_runstate(td, TD_REAPED);
1445                                 goto reaped;
1446                         }
1447                         if (WIFEXITED(status)) {
1448                                 if (WEXITSTATUS(status) && !td->error)
1449                                         td->error = WEXITSTATUS(status);
1450
1451                                 td_set_runstate(td, TD_REAPED);
1452                                 goto reaped;
1453                         }
1454                 }
1455
1456                 /*
1457                  * thread is not dead, continue
1458                  */
1459                 pending++;
1460                 continue;
1461 reaped:
1462                 (*nr_running)--;
1463                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1464                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1465                 if (!td->pid)
1466                         pending--;
1467
1468                 if (td->error)
1469                         exit_value++;
1470
1471                 done_secs += mtime_since_now(&td->epoch) / 1000;
1472         }
1473
1474         if (*nr_running == cputhreads && !pending && realthreads)
1475                 fio_terminate_threads(TERMINATE_ALL);
1476 }
1477
1478 /*
1479  * Main function for kicking off and reaping jobs, as needed.
1480  */
1481 static void run_threads(void)
1482 {
1483         struct thread_data *td;
1484         unsigned long spent;
1485         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1486
1487         if (fio_pin_memory())
1488                 return;
1489
1490         if (fio_gtod_offload && fio_start_gtod_thread())
1491                 return;
1492
1493         set_sig_handlers();
1494
1495         if (output_format == FIO_OUTPUT_NORMAL) {
1496                 log_info("Starting ");
1497                 if (nr_thread)
1498                         log_info("%d thread%s", nr_thread,
1499                                                 nr_thread > 1 ? "s" : "");
1500                 if (nr_process) {
1501                         if (nr_thread)
1502                                 log_info(" and ");
1503                         log_info("%d process%s", nr_process,
1504                                                 nr_process > 1 ? "es" : "");
1505                 }
1506                 log_info("\n");
1507                 fflush(stdout);
1508         }
1509
1510         todo = thread_number;
1511         nr_running = 0;
1512         nr_started = 0;
1513         m_rate = t_rate = 0;
1514
1515         for_each_td(td, i) {
1516                 print_status_init(td->thread_number - 1);
1517
1518                 if (!td->o.create_serialize)
1519                         continue;
1520
1521                 /*
1522                  * do file setup here so it happens sequentially,
1523                  * we don't want X number of threads getting their
1524                  * client data interspersed on disk
1525                  */
1526                 if (setup_files(td)) {
1527                         exit_value++;
1528                         if (td->error)
1529                                 log_err("fio: pid=%d, err=%d/%s\n",
1530                                         (int) td->pid, td->error, td->verror);
1531                         td_set_runstate(td, TD_REAPED);
1532                         todo--;
1533                 } else {
1534                         struct fio_file *f;
1535                         unsigned int j;
1536
1537                         /*
1538                          * for sharing to work, each job must always open
1539                          * its own files. so close them, if we opened them
1540                          * for creation
1541                          */
1542                         for_each_file(td, f, j) {
1543                                 if (fio_file_open(f))
1544                                         td_io_close_file(td, f);
1545                         }
1546                 }
1547         }
1548
1549         set_genesis_time();
1550
1551         while (todo) {
1552                 struct thread_data *map[REAL_MAX_JOBS];
1553                 struct timeval this_start;
1554                 int this_jobs = 0, left;
1555
1556                 /*
1557                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1558                  */
1559                 for_each_td(td, i) {
1560                         if (td->runstate != TD_NOT_CREATED)
1561                                 continue;
1562
1563                         /*
1564                          * never got a chance to start, killed by other
1565                          * thread for some reason
1566                          */
1567                         if (td->terminate) {
1568                                 todo--;
1569                                 continue;
1570                         }
1571
1572                         if (td->o.start_delay) {
1573                                 spent = mtime_since_genesis();
1574
1575                                 if (td->o.start_delay * 1000 > spent)
1576                                         continue;
1577                         }
1578
1579                         if (td->o.stonewall && (nr_started || nr_running)) {
1580                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1581                                                         td->o.name);
1582                                 break;
1583                         }
1584
1585                         init_disk_util(td);
1586
1587                         /*
1588                          * Set state to created. Thread will transition
1589                          * to TD_INITIALIZED when it's done setting up.
1590                          */
1591                         td_set_runstate(td, TD_CREATED);
1592                         map[this_jobs++] = td;
1593                         nr_started++;
1594
1595                         if (td->o.use_thread) {
1596                                 int ret;
1597
1598                                 dprint(FD_PROCESS, "will pthread_create\n");
1599                                 ret = pthread_create(&td->thread, NULL,
1600                                                         thread_main, td);
1601                                 if (ret) {
1602                                         log_err("pthread_create: %s\n",
1603                                                         strerror(ret));
1604                                         nr_started--;
1605                                         break;
1606                                 }
1607                                 ret = pthread_detach(td->thread);
1608                                 if (ret)
1609                                         log_err("pthread_detach: %s",
1610                                                         strerror(ret));
1611                         } else {
1612                                 pid_t pid;
1613                                 dprint(FD_PROCESS, "will fork\n");
1614                                 pid = fork();
1615                                 if (!pid) {
1616                                         int ret = fork_main(shm_id, i);
1617
1618                                         _exit(ret);
1619                                 } else if (i == fio_debug_jobno)
1620                                         *fio_debug_jobp = pid;
1621                         }
1622                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1623                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1624                                 log_err("fio: job startup hung? exiting.\n");
1625                                 fio_terminate_threads(TERMINATE_ALL);
1626                                 fio_abort = 1;
1627                                 nr_started--;
1628                                 break;
1629                         }
1630                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1631                 }
1632
1633                 /*
1634                  * Wait for the started threads to transition to
1635                  * TD_INITIALIZED.
1636                  */
1637                 fio_gettime(&this_start, NULL);
1638                 left = this_jobs;
1639                 while (left && !fio_abort) {
1640                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1641                                 break;
1642
1643                         usleep(100000);
1644
1645                         for (i = 0; i < this_jobs; i++) {
1646                                 td = map[i];
1647                                 if (!td)
1648                                         continue;
1649                                 if (td->runstate == TD_INITIALIZED) {
1650                                         map[i] = NULL;
1651                                         left--;
1652                                 } else if (td->runstate >= TD_EXITED) {
1653                                         map[i] = NULL;
1654                                         left--;
1655                                         todo--;
1656                                         nr_running++; /* work-around... */
1657                                 }
1658                         }
1659                 }
1660
1661                 if (left) {
1662                         log_err("fio: %d job%s failed to start\n", left,
1663                                         left > 1 ? "s" : "");
1664                         for (i = 0; i < this_jobs; i++) {
1665                                 td = map[i];
1666                                 if (!td)
1667                                         continue;
1668                                 kill(td->pid, SIGTERM);
1669                         }
1670                         break;
1671                 }
1672
1673                 /*
1674                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1675                  */
1676                 for_each_td(td, i) {
1677                         if (td->runstate != TD_INITIALIZED)
1678                                 continue;
1679
1680                         if (in_ramp_time(td))
1681                                 td_set_runstate(td, TD_RAMP);
1682                         else
1683                                 td_set_runstate(td, TD_RUNNING);
1684                         nr_running++;
1685                         nr_started--;
1686                         m_rate += ddir_rw_sum(td->o.ratemin);
1687                         t_rate += ddir_rw_sum(td->o.rate);
1688                         todo--;
1689                         fio_mutex_up(td->mutex);
1690                 }
1691
1692                 reap_threads(&nr_running, &t_rate, &m_rate);
1693
1694                 if (todo) {
1695                         if (is_backend)
1696                                 fio_server_idle_loop();
1697                         else
1698                                 usleep(100000);
1699                 }
1700         }
1701
1702         while (nr_running) {
1703                 reap_threads(&nr_running, &t_rate, &m_rate);
1704
1705                 if (is_backend)
1706                         fio_server_idle_loop();
1707                 else
1708                         usleep(10000);
1709         }
1710
1711         update_io_ticks();
1712         fio_unpin_memory();
1713 }
1714
1715 void wait_for_disk_thread_exit(void)
1716 {
1717         fio_mutex_down(disk_thread_mutex);
1718 }
1719
1720 static void free_disk_util(void)
1721 {
1722         disk_util_start_exit();
1723         wait_for_disk_thread_exit();
1724         disk_util_prune_entries();
1725 }
1726
1727 static void *disk_thread_main(void *data)
1728 {
1729         int ret = 0;
1730
1731         fio_mutex_up(startup_mutex);
1732
1733         while (threads && !ret) {
1734                 usleep(DISK_UTIL_MSEC * 1000);
1735                 if (!threads)
1736                         break;
1737                 ret = update_io_ticks();
1738
1739                 if (!is_backend)
1740                         print_thread_status();
1741         }
1742
1743         fio_mutex_up(disk_thread_mutex);
1744         return NULL;
1745 }
1746
1747 static int create_disk_util_thread(void)
1748 {
1749         int ret;
1750
1751         setup_disk_util();
1752
1753         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1754
1755         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1756         if (ret) {
1757                 fio_mutex_remove(disk_thread_mutex);
1758                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1759                 return 1;
1760         }
1761
1762         ret = pthread_detach(disk_util_thread);
1763         if (ret) {
1764                 fio_mutex_remove(disk_thread_mutex);
1765                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1766                 return 1;
1767         }
1768
1769         dprint(FD_MUTEX, "wait on startup_mutex\n");
1770         fio_mutex_down(startup_mutex);
1771         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1772         return 0;
1773 }
1774
1775 int fio_backend(void)
1776 {
1777         struct thread_data *td;
1778         int i;
1779
1780         if (exec_profile) {
1781                 if (load_profile(exec_profile))
1782                         return 1;
1783                 free(exec_profile);
1784                 exec_profile = NULL;
1785         }
1786         if (!thread_number)
1787                 return 0;
1788
1789         if (write_bw_log) {
1790                 setup_log(&agg_io_log[DDIR_READ], 0);
1791                 setup_log(&agg_io_log[DDIR_WRITE], 0);
1792                 setup_log(&agg_io_log[DDIR_TRIM], 0);
1793         }
1794
1795         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1796         if (startup_mutex == NULL)
1797                 return 1;
1798         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1799         if (writeout_mutex == NULL)
1800                 return 1;
1801
1802         set_genesis_time();
1803         create_disk_util_thread();
1804
1805         cgroup_list = smalloc(sizeof(*cgroup_list));
1806         INIT_FLIST_HEAD(cgroup_list);
1807
1808         run_threads();
1809
1810         if (!fio_abort) {
1811                 show_run_stats();
1812                 if (write_bw_log) {
1813                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1814                         __finish_log(agg_io_log[DDIR_WRITE],
1815                                         "agg-write_bw.log");
1816                         __finish_log(agg_io_log[DDIR_TRIM],
1817                                         "agg-write_bw.log");
1818                 }
1819         }
1820
1821         for_each_td(td, i)
1822                 fio_options_free(td);
1823
1824         free_disk_util();
1825         cgroup_kill(cgroup_list);
1826         sfree(cgroup_list);
1827         sfree(cgroup_mnt);
1828
1829         fio_mutex_remove(startup_mutex);
1830         fio_mutex_remove(writeout_mutex);
1831         fio_mutex_remove(disk_thread_mutex);
1832         return exit_value;
1833 }